Segmentation of organs at risk (OAR) in male pelvis is critical for planning prostate cancer radiotherapy. We are interested in femoral heads, rectum and bladder segmentation in magnetic resonance imaging (MRI) and computed tomography (CT) images in order to protect OARs during radiotherapy planning. The proposed methodology is based on superpixel algorithm in order to over-segment patient image by solving a local Eikonal function from initial seeds.
Afterwards, thesegmentation is obtained by computing a graph diffusion on a region adjacency graph (RAG) extracted from the over-segmentation thanks to some nodes labeled by the user. Superpixel segmentation is carried out slice-by-slice in 2D. Then, a RAG is constructed in 3D to obtain 3D OAR segmentation. The influence of the initial number of seeds on the segmentation is studied. The performances of the algorithm is evaluated and compared to 4 other methods.